Probabilistic Discriminative Dimensionality Reduction for Pose-Based Action Recognition

We examine the problem of classifying action sequences given a small set of examples for each type of action. Based on the presumption that human motion resides in a low dimensional space, we introduce a probabilistic dimensionality reduction model able to recover the structure of a low-dimensional manifold where all the involved actions reside. Requiring that sequences of the same action are placed apart from other sequences, we are able to achieve higher classification rates, with respect to other commonly used techniques, by performing the classification on this manifold. The main contribution is the introduction of a new model, based on Back-constrained GP-LVM which can be used for the efficient classification of sequences. We compare our method with the classification based on the Dynamic Time Warping distance and with the V-GPDS model, adapted for classification. Results are provided for sequences taken from two publicly available datasets which highlight different aspects of the method.